On Sample-Efficient Generalized Planning via Learned Transition Models

Nitin Gupta; Vishal Pallagani; John A. Aydin; Biplav Srivastava

cs.AIFeb 26, 2026

On Sample-Efficient Generalized Planning via Learned Transition Models

Authors: Nitin Gupta, Vishal Pallagani, John A. Aydin, Biplav Srivastava

Organizations: University of South Carolina

Abstract

Generalized planning studies the construction of solution strategies that generalize across families of planning problems sharing a common domain model, formally defined by a transition function $γ: S \times A \rightarrow S$ . Classical approaches achieve such generalization through symbolic abstractions and explicit reasoning over $γ$ . In contrast, recent Transformer-based planners, such as PlanGPT and Plansformer, largely cast generalized planning as direct action-sequence prediction, bypassing explicit transition modeling. While effective on in-distribution instances, these approaches typically require large datasets and model sizes, and often suffer from state drift in long-horizon settings due to the absence of explicit world-state evolution. In this work, we formulate generalized planning as a transition-model learning problem, in which a neural model explicitly approximates the successor-state function $\hatγ \approx γ$ and generates plans by rolling out symbolic state trajectories. Instead of predicting actions directly, the model autoregressively predicts intermediate world states, thereby learning the domain dynamics as an implicit world model. To study size-invariant generalization and sample efficiency, we systematically evaluate multiple state representations and neural architectures, including relational graph encodings. Our results show that learning explicit transition models yields higher out-of-distribution satisficing-plan success than direct action-sequence prediction in multiple domains, while achieving these gains with significantly fewer training instances and smaller models. This is an extended version of a short paper accepted at ICAPS 2026 under the same title.

Figures & tables

Figure 1: State-Centric Generalized Planning Pipeline. From a symbolic planning instance $\Pi$ , executable plans are generated using a learned transition model. (1) State Encoding: Symbolic state–goal pairs $(s_{t},g)$ are mapped to fixed-dimensional embeddings $\phi(s_{t})$ using either WL graph kernels or fixed-size factored vectors. (2) Transition Modeling: A parametric model (LSTM) or a non-parametric model (XGBoost) learns residual state transitions $\Delta_{t}$ to predict successor embeddings. (3) Neuro-Symbolic Plan Decoding: The predicted successor embedding $\hat{\phi}(s_{t+1})$ is matched against all valid symbolic successors $\mathrm{Succ}(s_{t})$ induced by $\gamma$ , and the nearest valid successor is selected to recover the executable action. This guarantees symbolic validity while enabling transition-model-based generalization.

		FD	Plansf.	PlanGPT ∗			SymT ${}^{\text{E}}$ ∗	SymT ${}^{\text{ED}}$ ∗			WL-LSTM		WL-XGB		FSF-LSTM		FSF-XGB
Domain	Split			greedy	appl.	regr.	greedy	greedy	appl.	regr.	state	delta	state	delta	state	delta	state	delta
Blocks	Val.	1.00	1.00 ±0.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00	0.00 ±0.00	0.44 ±0.16	0.67 ${}_{\scriptscriptstyle\pm\underline{0.00}}$	1.00 ±0.00	0.67 ${}_{\scriptscriptstyle\pm\underline{0.00}}$	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00
	Interp.	1.00	1.00 ±0.00	0.56 ±0.16	0.56 ±0.16	0.00 ±0.00	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00	0.67 ±0.27	0.89 ${}_{\scriptscriptstyle\pm\underline{0.16}}$	1.00 ±0.00	1.00 ±0.00	0.11 ±0.16	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00
	Extrap.	0.60	0.10 ±0.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00	0.05 ±0.07	0.07 ±0.02	0.13 ±0.05	0.00 ±0.00	0.10 ±0.07	0.15 ±0.07	0.25 ${}_{\scriptscriptstyle\pm\underline{0.00}}$	0.50 ±0.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00
Gripper	Val.	1.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00	1.00 ±0.00	0.17 ±0.24	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00	0.67 ${}_{\scriptscriptstyle\pm\underline{0.47}}$	0.00 ±0.00	0.00 ±0.00	0.17 ±0.24	0.50 ±0.00	0.00 ±0.00	0.00 ±0.00
	Interp.	1.00	0.00 ±0.00	0.00 ±0.00	0.44 ±0.16	0.00 ±0.00	0.89 ${}_{\scriptscriptstyle\pm\underline{0.16}}$	0.67 ±0.00	1.00 ±0.00	1.00 ±0.00	1.00 ±0.00	0.67 ±0.47	0.00 ±0.00	0.00 ±0.00	0.44 ±0.31	0.22 ±0.31	0.67 ±0.00	0.00 ±0.00

Table 1: Coverage rates (%) across all configurations. Values are reported as

Mean_{\pm Std}

. Best result per row (over all non-FD configurations) is bolded , second best is underlined . Light gray columns denote our state-centric implementations. FD = Fast Downward (60s timeout). ∗ Results from Fritzsche et al. (2026) .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Symbol	Meaning
Planning	$\Sigma=\langle S,A,\gamma\rangle$	State-transition system (set of states $S$ , actions $A$ , and transition function $\gamma$ )
	$S$	Set of all symbolic world states
	$A,\mathcal{A}$	Set of grounded actions (operators)
	$\gamma:S\times\mathcal{A}\rightarrow S$	Deterministic transition function
	$\mathcal{O}$	Finite object set in a planning instance
	$\mathcal{P}$	Predicate vocabulary

Appendix

Table 2: Summary of notation used throughout the paper.

Domain	FSF Dim. ( $D$ )	WL Dim. ( $D$ )	WL-LSTM Params	WL-XGBoost Nodes
Blocksworld	17	172	1,253,292	335,154
Gripper	46	302	1,552,942	128,310
VisitAll	121	138	1,174,922	679,060
Logistics	37	552	2,129,192	819,320

Appendix

Table 3: State representation dimensions (

D

) and model complexity.

Method	Parameters	Architecture	Ratio
Plansformer	$\sim$ 220M	CodeT5	$\sim$ 230 $\times$
PlanGPT	$\sim$ 125M	GPT-2	$\sim$ 130 $\times$
SymT ${}^{\text{ED}}$	$\sim$ 25–35M	Transformer	$\sim$ 30 $\times$
Ours (LSTM)	$\sim$ 1.1–2.1M	2-layer LSTM	1 $\times$
Ours (XGB)	$\sim$ 340–820K nodes	Gradient boosting	—

Appendix

Table 4: Model size comparison across all methods. SymT estimates are derived from architectural specifications in Fritzsche et al. (2026) (12-layer shared-weight encoder/decoder, hidden size 768); exact counts depend on domain-specific vocabulary embeddings, which are not reported in the SymT work.

Figure 2: Satisficing-plan success rates on the validation split across all domains, comparing PlanGPT, SymT, WL-based, and FSF baselines.

Figure 3: Satisficing-plan success rates on the interpolation split, comparing PlanGPT, SymT, WL-based, and FSF baselines for generalization to in-distribution problem instances.

Figure 4: Satisficing-plan success rates on the extrapolation split, comparing PlanGPT, SymT, WL-based, and FSF baselines for generalization in out-of-distribution problem instances.

Domain	Train	Validation	Interpolation	Extrapolation
Blocksworld (blocks)	9 (4, 6, 7)	3 (8)	3 (5)	20 (9–17)
Gripper (balls)	4 (2, 4, 6, 8)	2 (9, 10)	3 (3, 5, 7)	16 (12–42, even)
Logistics (goals)	12 (1, 3, 5)	4 (6)	9 (2, 4)	18 (7–15)
VisitAll (cells)	207 (1, 3, 4, 6, 10, 11, 12, 14, 16)	24 (18, 20)	37 (2, 5, 8, 9, 15)	219 (24–121)

Appendix

Table 5: Dataset statistics. Each cell shows count (complexity values) for the corresponding split. Complexity is measured in blocks, balls, goals, or cells depending on the domain. Training uses small instances; extrapolation tests generalization to significantly larger problems.

Domain	Slot Interpretation	Value Semantics
Blocksworld	Slot 0: Unused (constant 0)	—
Blocksworld	Slot $i$ ( $i>0$ ): Block $i$	$0$ = on table; $-1$ = held by gripper; $j>0$ = on block $j$
Gripper	Slot 0: Robot location	Room index where robot is located
	Slot $i$ (ball): Ball $i$ location	Room index if at room; $-j$ if carried by gripper $j$
	Slot $i$ (gripper): Gripper $i$ status	$0$ = free; $j>0$ = holding ball $j$
Logistics	Slot 0: Unused (constant 0)	—

Appendix

Table 6: Detailed FSF encoding semantics by domain.

Architecture
Parameter	LSTM	XGBoost
Hidden dimension	256	—
Number of layers	2	—
Tree depth	—	8
Training
Learning rate	$10^{-2}$	0.1

Appendix

Table 7: Complete hyperparameter settings for all models.

Model-Based Planning

Model-Based Planning uses explicit or learned models of how actions change states to simulate futures, compare plans, and choose actions.

See all 206 papers

Aug 11, 2025

Symmetry-Aware Transformer Training for Automated Planning

While transformers excel in many settings, their application in the field of automated planning is limited. Prior work like PlanGPT, a state-of-the-art decoder-only transformer, struggles with extrapolation from easy to hard planning problems. This in turn stems from problem symmetries: planning tasks can be represented with arbitrary variable names that carry no meaning beyond being identifiers. This causes a combinatorial explosion of equivalent representations that pure transformers cannot efficiently learn from. We propose a novel contrastive learning objective to make transformers symmetry-aware and thereby compensate for their lack of inductive bias. Combining this with architectural improvements, we show that transformers can be efficiently trained for either plan-generation or heuristic-prediction. Our results across multiple planning domains demonstrate that our symmetry-aware training effectively and efficiently addresses the limitations of PlanGPT.

Markus Fritzsche, Elliot Gestrin, Jendrik Seipp

Linköping University, Sweden

Contrastive Learning Transformer Symbolic Planning

Ask PDF

Feb 4, 2026

Textual Planning with Explicit Latent Transitions

Planning requires a transition model that predicts how each action changes the current state. When a large language model (LLM) plays this role, every next state is generated token by token, which makes searching over many possible futures slow and expensive. Existing alternatives either still query an LLM at every step or require a symbolic model of the domain. We propose EmbedPlan, a transition model built on frozen text embeddings: it embeds natural language descriptions of the state and the action with a frozen LLM, predicts the embedding of the next state with a lightweight learned network, and returns the closest real state. Because this network can be trained on top of any encoder, EmbedPlan also provides a controlled way to compare text representations for learning transitions. We evaluate it on 9 classical planning domains, under six settings that hold out progressively more of the data, from transitions to entire domains, and against baselines ranging from predicting no change to learning symbolic action rules. On planning problems seen during training, EmbedPlan almost always ranks the true next state among its top five guesses, still does so for most queries even when every observed state is a candidate, and retains 92-99% of its single-step accuracy when predicting several steps ahead from its own outputs. Given the same candidate states as GPT-5.4, it picks the true next state more often while taking about 0.17 ms per transition with cached embeddings. Accuracy is lower on unseen problems and near chance on unseen domains, and the controlled comparison traces this limit to the state representation rather than to the learned transition.

Eliezer Shlomi, Ido Levy, Eilam Shapira +6

Technion – Israel Institute of Technology, Haifa, Israel · IBM, Haifa, Israel.

Latent Dynamics Modeling LLM Planning Model-Based Planning

Ask PDF

Oct 19, 2026

AnyGoal: On World Model Generalization Beyond Expert Rollouts

The goal of a world model is to capture the dynamics of an environment. This has the great appeal that it allows agents to plan and simulate a course of actions with the help of a world model. For these uses to be reliable, however, the learned dynamics must behave well beyond the specific state-action combinations observed in expert rollouts. Despite this, the relationship between data coverage, dynamics generalization, and downstream planning performance remains underexplored. We study this relationship through the lens of compositional generalization: can a world model recombine familiar state and action factors into valid dynamics? To answer this, we introduce AnyGoal, a benchmark of goal-shifted task variants built from widely used world-model environments, together with an open-source library. Each task is designed to be within reach of a world model that has learned the underlying dynamics: individual state factors and actions remain within their training support, but their combinations are underrepresented in expert trajectories. Across multiple environments and world model families, we find that current models consistently exhibit failures under these recombinations. By examining dynamics and representation failures across model families, interaction regimes, and evaluation metrics, we study where current world models generalize and where challenges remain for goal-shifted prediction and planning. Together, this framework, benchmark, and analysis provide a basis for evaluating progress toward world models that generalize to new planning tasks.

Heejeong Nam, Divyat Mahajan, Quentin Le Lidec +2

Brown University · Cold Spring Harbor Laboratory · Mila, Université de Montréal +1

World Models World Model-Based Planning World Model Evaluation

Ask PDF

Abstract

Model-Based Planning

Explore similar work

Symmetry-Aware Transformer Training for Automated Planning

Textual Planning with Explicit Latent Transitions

AnyGoal: On World Model Generalization Beyond Expert Rollouts